949c97ff57e35bec6917c4aab4aca2ebaa165c32
[linux-2.6.git] / drivers / idle / i7300_idle.c
1 /*
2  * (C) Copyright 2008 Intel Corporation
3  * Authors:
4  * Andy Henroid <andrew.d.henroid@intel.com>
5  * Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
6  */
7
8 /*
9  * Save DIMM power on Intel 7300-based platforms when all CPUs/cores
10  * are idle, using the DIMM thermal throttling capability.
11  *
12  * This driver depends on the Intel integrated DMA controller (I/O AT).
13  * If the driver for I/O AT (drivers/dma/ioatdma*) is also enabled,
14  * this driver should work cooperatively.
15  */
16
17 /* #define DEBUG */
18
19 #include <linux/module.h>
20 #include <linux/pci.h>
21 #include <linux/sched.h>
22 #include <linux/notifier.h>
23 #include <linux/cpumask.h>
24 #include <linux/ktime.h>
25 #include <linux/delay.h>
26 #include <linux/debugfs.h>
27 #include <linux/stop_machine.h>
28 #include <linux/i7300_idle.h>
29
30 #include <asm/idle.h>
31
32 #include "../dma/ioatdma_hw.h"
33 #include "../dma/ioatdma_registers.h"
34
35 #define I7300_IDLE_DRIVER_VERSION       "1.55"
36 #define I7300_PRINT                     "i7300_idle:"
37
38 #define MAX_STOP_RETRIES        10
39
40 static int debug;
41 module_param_named(debug, debug, uint, 0644);
42 MODULE_PARM_DESC(debug, "Enable debug printks in this driver");
43
44 static int forceload;
45 module_param_named(forceload, forceload, uint, 0644);
46 MODULE_PARM_DESC(debug, "Enable driver testing on unvalidated i5000");
47
48 #define dprintk(fmt, arg...) \
49         do { if (debug) printk(KERN_INFO I7300_PRINT fmt, ##arg); } while (0)
50
51 /*
52  * Value to set THRTLOW to when initiating throttling
53  *  0 = No throttling
54  *  1 = Throttle when > 4 activations per eval window (Maximum throttling)
55  *  2 = Throttle when > 8 activations
56  *  168 = Throttle when > 672 activations (Minimum throttling)
57  */
58 #define MAX_THROTTLE_LOW_LIMIT          168
59 static uint throttle_low_limit = 1;
60 module_param_named(throttle_low_limit, throttle_low_limit, uint, 0644);
61 MODULE_PARM_DESC(throttle_low_limit,
62                 "Value for THRTLOWLM activation field "
63                 "(0 = disable throttle, 1 = Max throttle, 168 = Min throttle)");
64
65 /*
66  * simple invocation and duration statistics
67  */
68 static unsigned long total_starts;
69 static unsigned long total_us;
70
71 #ifdef DEBUG
72 static unsigned long past_skip;
73 #endif
74
75 static struct pci_dev *fbd_dev;
76
77 static spinlock_t i7300_idle_lock;
78 static int i7300_idle_active;
79
80 static u8 i7300_idle_thrtctl_saved;
81 static u8 i7300_idle_thrtlow_saved;
82 static u32 i7300_idle_mc_saved;
83
84 static cpumask_t idle_cpumask;
85 static ktime_t start_ktime;
86 static unsigned long avg_idle_us;
87
88 static struct dentry *debugfs_dir;
89
90 /* Begin: I/O AT Helper routines */
91
92 #define IOAT_CHANBASE(ioat_ctl, chan) (ioat_ctl + 0x80 + 0x80 * chan)
93 /* Snoop control (disable snoops when coherency is not important) */
94 #define IOAT_DESC_SADDR_SNP_CTL (1UL << 1)
95 #define IOAT_DESC_DADDR_SNP_CTL (1UL << 2)
96
97 static struct pci_dev *ioat_dev;
98 static struct ioat_dma_descriptor *ioat_desc; /* I/O AT desc & data (1 page) */
99 static unsigned long ioat_desc_phys;
100 static u8 *ioat_iomap; /* I/O AT memory-mapped control regs (aka CB_BAR) */
101 static u8 *ioat_chanbase;
102
103 /* Start I/O AT memory copy */
104 static int i7300_idle_ioat_start(void)
105 {
106         u32 err;
107         /* Clear error (due to circular descriptor pointer) */
108         err = readl(ioat_chanbase + IOAT_CHANERR_OFFSET);
109         if (err)
110                 writel(err, ioat_chanbase + IOAT_CHANERR_OFFSET);
111
112         writeb(IOAT_CHANCMD_START, ioat_chanbase + IOAT1_CHANCMD_OFFSET);
113         return 0;
114 }
115
116 /* Stop I/O AT memory copy */
117 static void i7300_idle_ioat_stop(void)
118 {
119         int i;
120         u64 sts;
121
122         for (i = 0; i < MAX_STOP_RETRIES; i++) {
123                 writeb(IOAT_CHANCMD_RESET,
124                         ioat_chanbase + IOAT1_CHANCMD_OFFSET);
125
126                 udelay(10);
127
128                 sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
129                         IOAT_CHANSTS_DMA_TRANSFER_STATUS;
130
131                 if (sts != IOAT_CHANSTS_DMA_TRANSFER_STATUS_ACTIVE)
132                         break;
133
134         }
135
136         if (i == MAX_STOP_RETRIES) {
137                 dprintk("failed to stop I/O AT after %d retries\n",
138                         MAX_STOP_RETRIES);
139         }
140 }
141
142 /* Test I/O AT by copying 1024 byte from 2k to 1k */
143 static int __init i7300_idle_ioat_selftest(u8 *ctl,
144                 struct ioat_dma_descriptor *desc, unsigned long desc_phys)
145 {
146         u64 chan_sts;
147
148         memset(desc, 0, 2048);
149         memset((u8 *) desc + 2048, 0xab, 1024);
150
151         desc[0].size = 1024;
152         desc[0].ctl = 0;
153         desc[0].src_addr = desc_phys + 2048;
154         desc[0].dst_addr = desc_phys + 1024;
155         desc[0].next = 0;
156
157         writeb(IOAT_CHANCMD_RESET, ioat_chanbase + IOAT1_CHANCMD_OFFSET);
158         writeb(IOAT_CHANCMD_START, ioat_chanbase + IOAT1_CHANCMD_OFFSET);
159
160         udelay(1000);
161
162         chan_sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
163                         IOAT_CHANSTS_DMA_TRANSFER_STATUS;
164
165         if (chan_sts != IOAT_CHANSTS_DMA_TRANSFER_STATUS_DONE) {
166                 /* Not complete, reset the channel */
167                 writeb(IOAT_CHANCMD_RESET,
168                        ioat_chanbase + IOAT1_CHANCMD_OFFSET);
169                 return -1;
170         }
171
172         if (*(u32 *) ((u8 *) desc + 3068) != 0xabababab ||
173             *(u32 *) ((u8 *) desc + 2044) != 0xabababab) {
174                 dprintk("Data values src 0x%x, dest 0x%x, memset 0x%x\n",
175                         *(u32 *) ((u8 *) desc + 2048),
176                         *(u32 *) ((u8 *) desc + 1024),
177                         *(u32 *) ((u8 *) desc + 3072));
178                 return -1;
179         }
180         return 0;
181 }
182
183 static struct device dummy_dma_dev = {
184         .init_name = "fallback device",
185         .coherent_dma_mask = DMA_BIT_MASK(64),
186         .dma_mask = &dummy_dma_dev.coherent_dma_mask,
187 };
188
189 /* Setup and initialize I/O AT */
190 /* This driver needs I/O AT as the throttling takes effect only when there is
191  * some memory activity. We use I/O AT to set up a dummy copy, while all CPUs
192  * go idle and memory is throttled.
193  */
194 static int __init i7300_idle_ioat_init(void)
195 {
196         u8 ver, chan_count, ioat_chan;
197         u16 chan_ctl;
198
199         ioat_iomap = (u8 *) ioremap_nocache(pci_resource_start(ioat_dev, 0),
200                                             pci_resource_len(ioat_dev, 0));
201
202         if (!ioat_iomap) {
203                 printk(KERN_ERR I7300_PRINT "failed to map I/O AT registers\n");
204                 goto err_ret;
205         }
206
207         ver = readb(ioat_iomap + IOAT_VER_OFFSET);
208         if (ver != IOAT_VER_1_2) {
209                 printk(KERN_ERR I7300_PRINT "unknown I/O AT version (%u.%u)\n",
210                         ver >> 4, ver & 0xf);
211                 goto err_unmap;
212         }
213
214         chan_count = readb(ioat_iomap + IOAT_CHANCNT_OFFSET);
215         if (!chan_count) {
216                 printk(KERN_ERR I7300_PRINT "unexpected # of I/O AT channels "
217                         "(%u)\n",
218                         chan_count);
219                 goto err_unmap;
220         }
221
222         ioat_chan = chan_count - 1;
223         ioat_chanbase = IOAT_CHANBASE(ioat_iomap, ioat_chan);
224
225         chan_ctl = readw(ioat_chanbase + IOAT_CHANCTRL_OFFSET);
226         if (chan_ctl & IOAT_CHANCTRL_CHANNEL_IN_USE) {
227                 printk(KERN_ERR I7300_PRINT "channel %d in use\n", ioat_chan);
228                 goto err_unmap;
229         }
230
231         writew(IOAT_CHANCTRL_CHANNEL_IN_USE,
232                 ioat_chanbase + IOAT_CHANCTRL_OFFSET);
233
234         ioat_desc = (struct ioat_dma_descriptor *)dma_alloc_coherent(
235                         &dummy_dma_dev, 4096,
236                         (dma_addr_t *)&ioat_desc_phys, GFP_KERNEL);
237         if (!ioat_desc) {
238                 printk(KERN_ERR I7300_PRINT "failed to allocate I/O AT desc\n");
239                 goto err_mark_unused;
240         }
241
242         writel(ioat_desc_phys & 0xffffffffUL,
243                ioat_chanbase + IOAT1_CHAINADDR_OFFSET_LOW);
244         writel(ioat_desc_phys >> 32,
245                ioat_chanbase + IOAT1_CHAINADDR_OFFSET_HIGH);
246
247         if (i7300_idle_ioat_selftest(ioat_iomap, ioat_desc, ioat_desc_phys)) {
248                 printk(KERN_ERR I7300_PRINT "I/O AT self-test failed\n");
249                 goto err_free;
250         }
251
252         /* Setup circular I/O AT descriptor chain */
253         ioat_desc[0].ctl = IOAT_DESC_SADDR_SNP_CTL | IOAT_DESC_DADDR_SNP_CTL;
254         ioat_desc[0].src_addr = ioat_desc_phys + 2048;
255         ioat_desc[0].dst_addr = ioat_desc_phys + 3072;
256         ioat_desc[0].size = 128;
257         ioat_desc[0].next = ioat_desc_phys + sizeof(struct ioat_dma_descriptor);
258
259         ioat_desc[1].ctl = ioat_desc[0].ctl;
260         ioat_desc[1].src_addr = ioat_desc[0].src_addr;
261         ioat_desc[1].dst_addr = ioat_desc[0].dst_addr;
262         ioat_desc[1].size = ioat_desc[0].size;
263         ioat_desc[1].next = ioat_desc_phys;
264
265         return 0;
266
267 err_free:
268         dma_free_coherent(&dummy_dma_dev, 4096, (void *)ioat_desc, 0);
269 err_mark_unused:
270         writew(0, ioat_chanbase + IOAT_CHANCTRL_OFFSET);
271 err_unmap:
272         iounmap(ioat_iomap);
273 err_ret:
274         return -ENODEV;
275 }
276
277 /* Cleanup I/O AT */
278 static void __exit i7300_idle_ioat_exit(void)
279 {
280         int i;
281         u64 chan_sts;
282
283         i7300_idle_ioat_stop();
284
285         /* Wait for a while for the channel to halt before releasing */
286         for (i = 0; i < MAX_STOP_RETRIES; i++) {
287                 writeb(IOAT_CHANCMD_RESET,
288                        ioat_chanbase + IOAT1_CHANCMD_OFFSET);
289
290                 chan_sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
291                         IOAT_CHANSTS_DMA_TRANSFER_STATUS;
292
293                 if (chan_sts != IOAT_CHANSTS_DMA_TRANSFER_STATUS_ACTIVE) {
294                         writew(0, ioat_chanbase + IOAT_CHANCTRL_OFFSET);
295                         break;
296                 }
297                 udelay(1000);
298         }
299
300         chan_sts = readq(ioat_chanbase + IOAT1_CHANSTS_OFFSET) &
301                         IOAT_CHANSTS_DMA_TRANSFER_STATUS;
302
303         /*
304          * We tried to reset multiple times. If IO A/T channel is still active
305          * flag an error and return without cleanup. Memory leak is better
306          * than random corruption in that extreme error situation.
307          */
308         if (chan_sts == IOAT_CHANSTS_DMA_TRANSFER_STATUS_ACTIVE) {
309                 printk(KERN_ERR I7300_PRINT "Unable to stop IO A/T channels."
310                         " Not freeing resources\n");
311                 return;
312         }
313
314         dma_free_coherent(&dummy_dma_dev, 4096, (void *)ioat_desc, 0);
315         iounmap(ioat_iomap);
316 }
317
318 /* End: I/O AT Helper routines */
319
320 #define DIMM_THRTLOW 0x64
321 #define DIMM_THRTCTL 0x67
322 #define DIMM_THRTCTL_THRMHUNT (1UL << 0)
323 #define DIMM_MC 0x40
324 #define DIMM_GTW_MODE (1UL << 17)
325 #define DIMM_GBLACT 0x60
326
327 /*
328  * Keep track of an exponential-decaying average of recent idle durations.
329  * The latest duration gets DURATION_WEIGHT_PCT percentage weight
330  * in this average, with the old average getting the remaining weight.
331  *
332  * High weights emphasize recent history, low weights include long history.
333  */
334 #define DURATION_WEIGHT_PCT 55
335
336 /*
337  * When the decaying average of recent durations or the predicted duration
338  * of the next timer interrupt is shorter than duration_threshold, the
339  * driver will decline to throttle.
340  */
341 #define DURATION_THRESHOLD_US 100
342
343
344 /* Store DIMM thermal throttle configuration */
345 static int i7300_idle_thrt_save(void)
346 {
347         u32 new_mc_val;
348         u8 gblactlm;
349
350         pci_read_config_byte(fbd_dev, DIMM_THRTCTL, &i7300_idle_thrtctl_saved);
351         pci_read_config_byte(fbd_dev, DIMM_THRTLOW, &i7300_idle_thrtlow_saved);
352         pci_read_config_dword(fbd_dev, DIMM_MC, &i7300_idle_mc_saved);
353         /*
354          * Make sure we have Global Throttling Window Mode set to have a
355          * "short" window. This (mostly) works around an issue where
356          * throttling persists until the end of the global throttling window
357          * size. On the tested system, this was resulting in a maximum of
358          * 64 ms to exit throttling (average 32 ms). The actual numbers
359          * depends on system frequencies. Setting the short window reduces
360          * this by a factor of 4096.
361          *
362          * We will only do this only if the system is set for
363          * unlimited-activations while in open-loop throttling (i.e., when
364          * Global Activation Throttle Limit is zero).
365          */
366         pci_read_config_byte(fbd_dev, DIMM_GBLACT, &gblactlm);
367         dprintk("thrtctl_saved = 0x%02x, thrtlow_saved = 0x%02x\n",
368                 i7300_idle_thrtctl_saved,
369                 i7300_idle_thrtlow_saved);
370         dprintk("mc_saved = 0x%08x, gblactlm = 0x%02x\n",
371                 i7300_idle_mc_saved,
372                 gblactlm);
373         if (gblactlm == 0) {
374                 new_mc_val = i7300_idle_mc_saved | DIMM_GTW_MODE;
375                 pci_write_config_dword(fbd_dev, DIMM_MC, new_mc_val);
376                 return 0;
377         } else {
378                 dprintk("could not set GTW_MODE = 1 (OLTT enabled)\n");
379                 return -ENODEV;
380         }
381 }
382
383 /* Restore DIMM thermal throttle configuration */
384 static void i7300_idle_thrt_restore(void)
385 {
386         pci_write_config_dword(fbd_dev, DIMM_MC, i7300_idle_mc_saved);
387         pci_write_config_byte(fbd_dev, DIMM_THRTLOW, i7300_idle_thrtlow_saved);
388         pci_write_config_byte(fbd_dev, DIMM_THRTCTL, i7300_idle_thrtctl_saved);
389 }
390
391 /* Enable DIMM thermal throttling */
392 static void i7300_idle_start(void)
393 {
394         u8 new_ctl;
395         u8 limit;
396
397         new_ctl = i7300_idle_thrtctl_saved & ~DIMM_THRTCTL_THRMHUNT;
398         pci_write_config_byte(fbd_dev, DIMM_THRTCTL, new_ctl);
399
400         limit = throttle_low_limit;
401         if (unlikely(limit > MAX_THROTTLE_LOW_LIMIT))
402                 limit = MAX_THROTTLE_LOW_LIMIT;
403
404         pci_write_config_byte(fbd_dev, DIMM_THRTLOW, limit);
405
406         new_ctl = i7300_idle_thrtctl_saved | DIMM_THRTCTL_THRMHUNT;
407         pci_write_config_byte(fbd_dev, DIMM_THRTCTL, new_ctl);
408 }
409
410 /* Disable DIMM thermal throttling */
411 static void i7300_idle_stop(void)
412 {
413         u8 new_ctl;
414         u8 got_ctl;
415
416         new_ctl = i7300_idle_thrtctl_saved & ~DIMM_THRTCTL_THRMHUNT;
417         pci_write_config_byte(fbd_dev, DIMM_THRTCTL, new_ctl);
418
419         pci_write_config_byte(fbd_dev, DIMM_THRTLOW, i7300_idle_thrtlow_saved);
420         pci_write_config_byte(fbd_dev, DIMM_THRTCTL, i7300_idle_thrtctl_saved);
421         pci_read_config_byte(fbd_dev, DIMM_THRTCTL, &got_ctl);
422         WARN_ON_ONCE(got_ctl != i7300_idle_thrtctl_saved);
423 }
424
425
426 /*
427  * i7300_avg_duration_check()
428  * return 0 if the decaying average of recent idle durations is
429  * more than DURATION_THRESHOLD_US
430  */
431 static int i7300_avg_duration_check(void)
432 {
433         if (avg_idle_us >= DURATION_THRESHOLD_US)
434                 return 0;
435
436 #ifdef DEBUG
437         past_skip++;
438 #endif
439         return 1;
440 }
441
442 /* Idle notifier to look at idle CPUs */
443 static int i7300_idle_notifier(struct notifier_block *nb, unsigned long val,
444                                 void *data)
445 {
446         unsigned long flags;
447         ktime_t now_ktime;
448         static ktime_t idle_begin_time;
449         static int time_init = 1;
450
451         if (!throttle_low_limit)
452                 return 0;
453
454         if (unlikely(time_init)) {
455                 time_init = 0;
456                 idle_begin_time = ktime_get();
457         }
458
459         spin_lock_irqsave(&i7300_idle_lock, flags);
460         if (val == IDLE_START) {
461
462                 cpu_set(smp_processor_id(), idle_cpumask);
463
464                 if (cpus_weight(idle_cpumask) != num_online_cpus())
465                         goto end;
466
467                 now_ktime = ktime_get();
468                 idle_begin_time = now_ktime;
469
470                 if (i7300_avg_duration_check())
471                         goto end;
472
473                 i7300_idle_active = 1;
474                 total_starts++;
475                 start_ktime = now_ktime;
476
477                 i7300_idle_start();
478                 i7300_idle_ioat_start();
479
480         } else if (val == IDLE_END) {
481                 cpu_clear(smp_processor_id(), idle_cpumask);
482                 if (cpus_weight(idle_cpumask) == (num_online_cpus() - 1)) {
483                         /* First CPU coming out of idle */
484                         u64 idle_duration_us;
485
486                         now_ktime = ktime_get();
487
488                         idle_duration_us = ktime_to_us(ktime_sub
489                                                 (now_ktime, idle_begin_time));
490
491                         avg_idle_us =
492                                 ((100 - DURATION_WEIGHT_PCT) * avg_idle_us +
493                                  DURATION_WEIGHT_PCT * idle_duration_us) / 100;
494
495                         if (i7300_idle_active) {
496                                 ktime_t idle_ktime;
497
498                                 idle_ktime = ktime_sub(now_ktime, start_ktime);
499                                 total_us += ktime_to_us(idle_ktime);
500
501                                 i7300_idle_ioat_stop();
502                                 i7300_idle_stop();
503                                 i7300_idle_active = 0;
504                         }
505                 }
506         }
507 end:
508         spin_unlock_irqrestore(&i7300_idle_lock, flags);
509         return 0;
510 }
511
512 static struct notifier_block i7300_idle_nb = {
513         .notifier_call = i7300_idle_notifier,
514 };
515
516 MODULE_DEVICE_TABLE(pci, pci_tbl);
517
518 int stats_open_generic(struct inode *inode, struct file *fp)
519 {
520         fp->private_data = inode->i_private;
521         return 0;
522 }
523
524 static ssize_t stats_read_ul(struct file *fp, char __user *ubuf, size_t count,
525                                 loff_t *off)
526 {
527         unsigned long *p = fp->private_data;
528         char buf[32];
529         int len;
530
531         len = snprintf(buf, 32, "%lu\n", *p);
532         return simple_read_from_buffer(ubuf, count, off, buf, len);
533 }
534
535 static const struct file_operations idle_fops = {
536         .open   = stats_open_generic,
537         .read   = stats_read_ul,
538 };
539
540 struct debugfs_file_info {
541         void *ptr;
542         char name[32];
543         struct dentry *file;
544 } debugfs_file_list[] = {
545                                 {&total_starts, "total_starts", NULL},
546                                 {&total_us, "total_us", NULL},
547 #ifdef DEBUG
548                                 {&past_skip, "past_skip", NULL},
549 #endif
550                                 {NULL, "", NULL}
551                         };
552
553 static int __init i7300_idle_init(void)
554 {
555         spin_lock_init(&i7300_idle_lock);
556         cpus_clear(idle_cpumask);
557         total_us = 0;
558
559         if (i7300_idle_platform_probe(&fbd_dev, &ioat_dev, forceload))
560                 return -ENODEV;
561
562         if (i7300_idle_thrt_save())
563                 return -ENODEV;
564
565         if (i7300_idle_ioat_init())
566                 return -ENODEV;
567
568         debugfs_dir = debugfs_create_dir("i7300_idle", NULL);
569         if (debugfs_dir) {
570                 int i = 0;
571
572                 while (debugfs_file_list[i].ptr != NULL) {
573                         debugfs_file_list[i].file = debugfs_create_file(
574                                         debugfs_file_list[i].name,
575                                         S_IRUSR,
576                                         debugfs_dir,
577                                         debugfs_file_list[i].ptr,
578                                         &idle_fops);
579                         i++;
580                 }
581         }
582
583         idle_notifier_register(&i7300_idle_nb);
584
585         printk(KERN_INFO "i7300_idle: loaded v%s\n", I7300_IDLE_DRIVER_VERSION);
586         return 0;
587 }
588
589 static void __exit i7300_idle_exit(void)
590 {
591         idle_notifier_unregister(&i7300_idle_nb);
592
593         if (debugfs_dir) {
594                 int i = 0;
595
596                 while (debugfs_file_list[i].file != NULL) {
597                         debugfs_remove(debugfs_file_list[i].file);
598                         i++;
599                 }
600
601                 debugfs_remove(debugfs_dir);
602         }
603         i7300_idle_thrt_restore();
604         i7300_idle_ioat_exit();
605 }
606
607 module_init(i7300_idle_init);
608 module_exit(i7300_idle_exit);
609
610 MODULE_AUTHOR("Andy Henroid <andrew.d.henroid@intel.com>");
611 MODULE_DESCRIPTION("Intel Chipset DIMM Idle Power Saving Driver v"
612                         I7300_IDLE_DRIVER_VERSION);
613 MODULE_LICENSE("GPL");